Smart Energy Management

• Leverages IT & communication for real-time applications — a two-way exchange of information (vs the one-way traditional grid).
• Self-healing, adaptive, resilient and sustainable.
• Traditional → smart: few large plants → many small producers; centralized → decentralized market; top-to-bottom → bidirectional distribution; passive consumer → active participant.

Two types of AI in smart grids

The grid-level AI tasks

Two ways to monitor consumption

Context variables impacting home energy load (⭐ feeds Q23a)

• Types of appliances being used
• User's habits — derived via Indoor Localization (Lecture 5) and Human Activity Recognition; personalization is crucial; the profile is also depicted by appliance ON/OFF usage
• External weather conditions & seasonal variations (e.g., summer A/C), plus time and day

Two energy-conservation strategies (combinable)

Pipeline: smart-home sensor data → HAR → activities + user profile + weather/time → energy management algorithms → recommendations or automatic intervention.

This is the exact subject of the open-ended Q23a: the role of short-term load forecasting and the relevant context variables.

The "ingredients"

• Setting: a smart home with multiple appliances; power measured at regular intervals (minutes, 15 min, hours).
• Observation horizon: the historical time series available to the model — how much past data is used as input.
• Forecast horizon: the future time window to predict (minutes → hours → days ahead).
• Granularity: appliance level, device-category level (HVAC, lighting), or whole-home.
• Inputs: past consumption (observation horizon) + optional contextual variables (time of day, weather, occupancy patterns).

Two deep approaches

Q23b asks exactly this: which appliance categories can be re-scheduled, and how RL-based re-scheduling works.

⭐ Categories of home appliances — Q23b

Short-term load forecasting tells when energy is cheap; the challenge is understanding whether and when usage can actually be deferred: identify appliances consuming extra energy → predict when usage minimizes cost → schedule while minimizing user's discomfort.

Reinforcement Learning in a nutshell

An agent interacts with an environment that provides numeric reward signals; the goal is learning to take actions that maximize long-term reward.

• "Classic" RL (Q-learning): estimate Q(s,a) — the expected reward of action a in state s; pick the action maximizing it; update the estimate after observing the actual reward.
• Deep RL: a neural network implicitly learns Q(s,a) by maximizing collected reward.
• Exploration vs exploitation: at the start the agent explores (random actions, no policy yet); later it exploits the learned policy. Balance via ε-greedy: explore with probability ε (high at start, decays over time), exploit with 1−ε.

⭐ RL for energy management — Q23b

• Estimating comfort in the reward: average Waiting Time Rate (WTR) of appliance operations (users prefer short delays → minimizing WTR increases comfort); environmental conditions (temperature, humidity, air quality); emotion/stress recognition tools.
• Exploration problem in smart homes: random actions may hurt the user's comfort → pre-train the decision module in a controlled environment, then fine-tune in the target home with transfer learning.

P(t) = total load at time t · Pᵢ = load of the i-th appliance · e(t) = noise term
Goal: observing only P(t), estimate each Pᵢ

Load patterns: each appliance has a characteristic signature

The "classic" pipeline

• Event detection = segmentation by spotting consumption changes attributable to appliance state changes: steady-state changes (variations within a stable state) vs transient changes (specific transitions between states). (Echoes dynamic segmentation from Lecture 3.)
• Metric — MAE: for each appliance, minimize the mean absolute error between estimated and true consumption. It's a regression problem with multiple outputs.

Why deep learning became the NILM standard

• Learns features directly from raw data — removes the event-detection problem.
• Learns complex load signatures and temporal dependencies between appliances.
• More robust to noise and scalable.

Architectures

• Denoising AutoEncoder (DAE): "denoising" = extracting a clean signal from a noisy one. The encoder takes the aggregated consumption window → latent space → the decoder maps it to the power consumption of each appliance. (Same autoencoder machinery as Lecture 3's anomaly detection — different target.)
• Transformer architectures: 1D conv + positional embedding → transformer layers (Q/K/V scaled dot-product attention) → deconv/linear decoding. (Direct application of Lecture 4!)

Single-appliance vs multi-output models

How to train? The labeling challenge

• NILM is a supervised regression task — the ground truth is each appliance's consumption. But homes differ in number/type of appliances, so generalizing is hard.
• Submetering: a plug sensor per appliance to collect ground truth — costly and unrealistic (once bought, why remove them?).
• Manual annotation (turning appliances on/off) — may be prohibitive.
• Simulation: simulate per-device consumption and aggregate — but meta-information for all devices is hard, and synthetic data may be unrealistic.
• Transfer learning: adapt labeled source domains to an unlabeled target home, with a domain-adaptation loss learning domain-invariant features. Limits: source/target domains and their data distributions should be similar.

Countermeasures

Example privacy-preserving strategy

• Goals: compute a consumer's total consumption over a period (billing) and the regional total at an instant — while avoiding revealing an individual's instantaneous measurements.
• How: perturb consumption data with Laplacian noise (differential privacy) — the masked daily profile no longer exposes appliance usage events.

9 MCQs + 2 open-ended. Q1 is the simulation's Q8 verbatim; open-ended Q10 is the simulation's Q23 verbatim.